Everyone is arguing model vs harness right now.
I run agents all week. When a job fails, it is almost never the model. It is missing memory, missing gates, or no failure ledger.
Builders: what broke your last agent run, the model or the setup?
Claude agents no longer resend the whole conversation every turn. Faster, yes. Cheaper, no.
Hot take: most AI automations don't fail on latency. They fail on cost and reliability.
If you could fix one for your agents tomorrow, which one?
It's working very well for me despite I'm using it for both coding and video editing at a large scale, it all depends on how you are managing your sessions/tasks.
A quick thing you can try immediately, remove MCP servers for things you only use on your computer and use their CLI instead.
Dot by OpenAI has a feature almost nobody is talking about.
Right now you get one Dot, and you can't change its model.
But ask it to create specialists, and that's where it gets interesting. Dot can run up to 6 of them at once, and each one can use any model and any reasoning level you ask for, all the way up to GPT 6 Astra Ultra.
How to do it:
Simply ask it to create a sub agent for a specific task dedining model and its reasoning level.
That simple.
One Dot, directing a whole team running at the top reasoning level.